A GMPLS/OBS Network Architecture EnablingQoS-aware End-to-End Burst Transport

Pedro Pedroso∗, Mirosław Klinkowski†, Jordi Perelló∗, Salvatore Spadaro∗, Davide Careglio∗, Josep Solé-Pareta∗∗CCABA, Universitat Politècnica de Catalunya (UPC), Barcelona, Spain, e-mail: ppedroso@ac.upc.edu†National Institute of Telecommunications (NIT), Wrocław, Poland, e-mail: mklinkow@itl.waw.pl

Abstract—This paper introduces a Generalized Multi-ProtocolLabel Switching (GMPLS)-enabled Optical Burst Switched(OBS) network architecture featuring end-to-end QoS-awareburst transport services. This is achieved by setting up burstLabel Switched Paths (LSPs) properly dimensioned to match spe-cific burst drop probability requirements. These burst LSPs areused for specific guaranteed QoS levels, whereas the remainingnetwork capacity can be left for best-effort burst support. Aimingto ensure the requested burst drop probability figures even underbursty traffic patterns, burst LSPs’ performance is continuouslymonitored. Therefore, GMPLS-driven capacity reconfigurationscan be dynamically triggered whether unfavorable networkconditions are detected. Through the paper, the GMPLS/OBSarchitecture is firstly detailed, followed by the presentation of theoptimized methods used for the initial burst LSP dimensioning.The successful network performance is finally illustrated bysimulations on several network scenarios.

I. INTRODUCTION AND MOTIVATION

Paving the way to the future Internet, all-optical networks havebroken the limitations of Optical/Electronic/Optical (OEO)conversions in the network, which cannot match the transmis-sion rates offered by recent advances in Dense WavelengthDivision Multiplexing (DWDM) technology [1].

In this context, the Optical Burst Switching (OBS) paradigmleverages on the statistical multiplexing of optical data planeresources to enable sub-wavelength switching in the net-work [2]. This is achieved by aggregating the packets ar-riving at ingress nodes into bursts to be sent individuallythrough a bufferless optical network. With aims to minimizethe overhead introduced along the burst signaling process,OBS networks typically rely on one-way resource reservation.However, such on-the-fly resource reservation does not ensuresuccessful burst delivery, as they could encounter contention atintermediate nodes and become lost. Looking at the literature,a large number of techniques have been proposed to reduceburst losses in OBS networks, mostly focused on providingintelligence to the optical layer. Examples of such techniquesare, amongst many others, link congestion prediction [3][4],wavelength selection considering the streamline effect [5],feedback mechanisms to control network congestion and con-nection admission [6][7] or QoS-aware deflection routing [8].Nevertheless, none of them guarantee specific QoS figures asthe offered load to the network increases.

Conversely, Optical Circuit Switching [9] networks al-low end-to-end QoS compliant optical circuits betweensource/destination nodes. For the sake of flexibility, thesecircuits can be dynamically provisioned by a control plane

governing the underlying optical devices. To support thesetasks, the IETF has defined Generalized Multi-Protocol LabelSwitching (GMPLS) as a set of protocols covering the requiredsignaling, routing and management functionalities [10]. Nev-ertheless, GMPLS/OCS still provides very coarse granularity(a whole wavelength), being sub-wavelength switching onlyfeasible through electrical grooming.

In this paper we introduce a GMPLS-enabled OBS networkfeaturing QoS-guaranteed transport services. The rationalebehind our proposal is to move the overall network intelligenceto the GMPLS control plane, being the OBS layer only respon-sible for local contention resolution. To this end, an extendedGMPLS control plane is placed on top of the OBS layer,providing this one with burst LSPs provisioning capabilities.These burst LSPs are tailored to specific QoS requirements,guaranteeing the requested service levels.

The contribution of this work is three-fold. First of all, wedefine the GMPLS/OBS network architecture, detailing the be-havior of its control and transport planes (Section II). Second,we focus on the offline burst LSP configuration problem in theGMPLS/OBS network by proposing a Mixed Integer LinearProgramming (MILP) model (Section III). Third, we considersituations where unexpected traffic peaks or highly dynamicpatterns may appear and negatively impact on the providedQoS levels. In this context, we propose burst LSP performancemonitoring mechanisms able detect undesired traffic surgesand trigger burst LSP reconfiguration procedures when needed(Section IV). The performance of these contributions arefinally evaluated through simulations on different networkscenarios (Section V).

II. GMPLS/OBS NETWORK ARCHITECTURE

In the proposed GMPLS/OBS network architecture, the ex-tended GMPLS control plane lies on top as the actual OBSnetwork controller, setting up, maintaining, reconfiguring andtearing down burst LSPs (hereafter simply referred as LSPs)according to the client traffic demands and QoS requirements.While the OBS control layer must share the same physicalnetwork as the OBS data plane due to the strict time relationbetween Burst Control Packets (BCPs) and optical bursts, theGMPLS control plane can follow a different topology than thelatter, even supported on a separated network.

As mentioned before, the main benefit of OBS is thestatistical multiplexing, better exploiting those underlying dataplane resources. However, due to the lack of optical buffering,

Fig. 1. Virtual topologies in GMPLS/OBS network.

random burst losses can be experienced in OBS due to con-tention. In our GMPLS/OBS network architecture, the burstloss probability (BLP) is controlled by tightly dimensioningthose LSPs supporting burst data traffic.

In particular, a virtual topology (VT) of LSPs is deployedfor each offered QoS level, where all LSPs share a largeenough number of wavelengths per link so that the requestedend-to-end QoS can be provided. Note that wavelength shar-ing, which fosters the statistical multiplexing of networkresources, is limited to only those LSPs with the same QoSlevel. Otherwise, OBS nodes would have to be providedwith complex scheduling algorithms to enable QoS leveldifferentiation over the same set of output wavelengths.

In Fig. 1, an example of a GMPLS/OBS network configuredwith two virtual topologies and enabling two QoS levels isillustrated. On the right hand side of the figure, we can observethe GMPLS controller of node 2, in charge of configuringthe forwarding table of the OBS controller that configures theoptical switch by means of the scheduler. Four LSP crossingnode 2 are currently established in the network, two per eachvirtual topology. Four entries are hence configured in theforwarding table. While all LSPs use the same output port 3,only those LSPs with the same QoS share output wavelengths;bursts belonging to LSP1 and LSP2 can leave node 2 usingeither λ1, λ2, λ3, or λ6 while those belonging to LSP3 andLSP4 only λ4, λ5, or λ7.

For the sake of simplicity, only two service classes areconsidered in this work, namely, a High-Priority (HP) withguaranteed QoS level and a Best Effort (BE) class. Therefore,only a single virtual topology of LSPs for the HP class has tobe dimensioned in the network, restricting the BE class to usethe spare network capacity. Note, however, that multiple VTscan be dimensioned with the model in Section III.

This GMPLS/OBS architecture aims at keeping the OBSlayer as fast as possible, as only simple local decisions arerequired (e.g., select a wavelength among a set of pre-selectedones). In fact, virtual topology set up and reconfiguration

Fig. 2. Proposed extensions to the RSVP-TE Unnumbered interface ID sub-object for burst LSP signaling.

actions (e.g., due to undesired link congestion or link failures)are moved to the GMPLS-enabled control plane. To thisgoal, the standard GMPLS protocol set has to be extendedaccordingly. This section introduces the extensions to GMPLSRSVP-TE protocol required for the setup of the QoS-awareLSPs resulting from the dimensioning models in section III.Moreover, section IV presents the procedures at the GMPLSnode to adjust their capacity dynamically upon congestion.Due to the lack of space, however, we leave for future workthose procedures and GMPLS protocol extensions involved inonline LSP provisioning (i.e., LSPs supporting new demandsnot contemplated in the original traffic matrix).

In standard RSVP-TE [11], explicit routing is achieved bymeans of an EXPLICIT_ROUTE object in Path messages.This object encapsulates a list of sub-objects determining thenodes and links along the explicit route. In case of unnumberedlinks [12], Unnumbered interface ID sub-objects contain, foreach traversed node, the router IP address (Router ID) and theidentifier of the interface associated to the desired output link(e.g., wavelength). This allows only one wavelength per hopto be allocated with standard RSVP-TE. As mentioned before,however, burst LSPs may bundle several wavelengths per hop.Fig. 2 presents the extensions to Unnumbered interface ID sub-object for burst LSP signaling. In this case, the Router ID fieldis followed by the ID’s of the interfaces associated to eachwavelength to be allocated for the LSP on the downstreamlink. Departing from the network characteristics and the statictraffic matrix, the following section presents the dimensioningmodel to obtain the explicit routes and wavelengths for allLSPs composing the targeted set of VTs.

III. MILP MODEL FOR LSP DIMENSIONING

In order to guarantee certain level of QoS in terms of burstlosses, wavelength resources have to be dimensioned properly.In this section, we address the problem of the VT design thatconcerns the establishment of explicit LSPs (also referred aspaths in this section) and the allocation of wavelengths innetwork links to support connections with QoS guarantees.More specifically, we are looking for such network routingthat for given set of (long-term) traffic demands and end-to-end requirements on the burst loss rate minimizes theoverall number of allocated wavelengths (i.e., the wavelengthusage) in the network. To treat the problem of absoluteQoS guarantees analytically, we employ the non-reducedload approximation [13] of a common OBS network loss

model [14]. Modelling assumptions are then represented asa set of constraints in a Mixed Integer Liner Programming(MILP) formulation.

Due to the space limitations, the main modelling steps arebriefly presented in this paper. For more details as well assome heuristics for the VT design problem we refer to [15].

A. Notation

We use G = (V, E) to denote the graph of an OBS network;the set of nodes is denoted as V , and the set of unidirectionallinks is denoted as E . Link e ∈ E comprises We wavelengths.

Let P denote the set of predefined candidate LSPs betweensource s and termination t nodes, s, t ∈ V , and s 6= t. Eachpath p ∈ P is identified with a subset p ⊆ E . Adequately,subset Pe ⊆ P identifies all paths that go through link e. Letδ = maxδp : p ∈ P be the length of the longest path in thenetwork, where δp is the length (in hops) of path p.

Let D denote the set of demands with QoS guaran-tees, where each demand corresponds to a pair of source-termination nodes. For each demand d ∈ D, hd ∈ R+ denotesthe volume of traffic; for convenience, hp = hd for p ∈ Pd.

Let Pd ⊆ P denote the set of candidate LSPs supportingdemand d; P =

⋃d∈DPd. Each subset Pd comprises a (small)

number of paths, e.g., k shortest paths, and a burst can followone of them.

B. Modelling Assumptions

1) Routing: The network applies source-based routing. Theselection of path p from set Pd is performed according to adecision variable xp (also referred to as the routing variable).We assume unsplittable routing, in particular, a burst flow isrouted over path p iff xp = 1 and there is only one pathp ∈ Pd such that xp = 1. Accordingly, traffic ρp offered topath p ∈ Pd is calculated as ρp = xphd.

2) Burst Losses: Due to the complexity of the Erlang fixed-point computation in the common OBS network loss model[14], we assume a simplified model based on the non-reducedload calculation [13]. In this model, to estimate traffic loadρe offered to link e, we add up the traffic load ρp offered toeach path p ∈ P that crosses this link: ρe =

∑p∈P:p3eρp =∑

p∈P:p3exphp, e ∈ E . The use of such approximation isjustified by its accuracy, particularly under low overall burstlosses (below 10−2) [13].

Moreover, we take the common assumption in the literatureof i.e.d. burst arrivals, i.i.d. burst durations, together withthe assumption of the full wavelength conversion capabilityin network nodes. Accordingly, the Erlang B-loss formulaB(ρ, w) is used to model the probability Be that a burst islost in link e.

3) Burst Loss Guarantees: We assume each demand be-longing to a QoS class has the same end-to-end (e2e) burstloss probability Be2e requirements. To meet the goal of thee2e QoS for each demand d ∈ D, we assume that at eachlink the burst losses are kept below certain level Blink, i.e.,Be ≤ Blink,∀e ∈ E . For the rest of the paper, we considerBlink fixed, the same for each link, and determined according

to Blink = 1−(1−Be2e

)1/δ. This model is a common model

frequently used to assure QoS guarantees in loss networksand it is also applicable under unsplittable source routing inOBS [15].

4) Wavelength Allocation: We consider each QoS class hasa number of wavelengths allocated in network links which arenot shared with other QoS classes. Although, in this paper,we focus on a single QoS class, still the restricted approachallows to extend the model easily to the scenario with multipleQoS classes.

The last modelling step is to define a dimensioning functionFe (·) which for given traffic load ρe determines the minimumnumber of wavelengths to be allocated in link e so thatto satisfy the blocking Blink requirements. Such a functionis given by a discrete (discontinuous, step-increasing) linkdimensioning function Fe (ρe) =

⌈B−1(ρe, Blink)

⌉, where

B−1(ρe, Blink) is the inverse of the Erlang B Loss formulaextended to the real domain [16], and d·e is the ceilingfunction.

C. Problem Formulation

It is convenient to define aw as the maximal load supportedby w wavelengths given target blocking probability Blink, i.e.,aw = B−1(w,Blink). Although there is no close formulato calculate B−1, still we can use a line search method(see e.g., [17]) to find the root ρ∗ of function f(ρ) =Blink − B(ρ, w) so that to approximate the value of aw byaw = ρ∗ for each w ≤ max We : e ∈ E. Also, we introducea segmentation on load segments: bw = aw − aw−1, w =1..max We : e ∈ E.

Finally, we substitute Fe (·) with its piecewise linear approx-imation, Fe(ρe) = min w : aw ≥ ρe, which further allowsus to express the dimensioning function by means of a 0-1integer programming (IP) formulation [13]. This formulationmakes use of a set of binary variables uwe : e ∈ E , w =1..We; uwe is active iff w or more wavelengths are allocatedin link e.

Our VT design problem can be formulated as a MILPproblem:

minimize∑

e

∑wuwe (MILP)

subject to∑p∈Pd

xp = 1, ∀d ∈ D, (1a)∑p∈P:p3ehpxp − ρe = 0, ∀e ∈ E , (1b)

ρe ≤ aWe , ∀e ∈ E , (1c)∑w=1..We

uwe bw − ρe ≥ 0, ∀e ∈ E , (1d)

uwe − uw+1e ≥ 0, ∀e ∈ E , w = 1..We − 1, (1e)

uwe ∈ 0, 1 , ∀e ∈ E , w = 1..We, (1f)

x ∈ 0, 1|P|, ρ ∈ R|E|+ , (1g)

where ρe is an auxiliary variable representing load in link e.The objective of the optimization problem is to minimize

the total number of wavelengths utilized in the network. (1a)

are the routing constraints. (1b) are auxiliary constraints ofthe non-reduced load calculation. (1c) are the link capacityconstraints. (1d) and (1e) result from the 0-1 representationof function Fe (·). In particular, the number of wavelengthsin link e should be such that the maximum traffic load it cansupport (calculated as the sum of active load segments bw)is greater or equal to offered traffic load ρe. Besides, (1e) areordering constraints, i.e., if w wavelengths are utilized so w−1wavelengths are utilized as well. Finally, (1f) and (1g) are thevariable range constraints.

Note that (MILP) is a variant of the well-known discrete costmulticommodity flow problem (DCMCF) which is a difficultproblem [18]. Still, performed experiments show that a goodsub-optimal solution (optimality gap below 2%) can be foundin reasonable time (from several to some hundreds of seconds)for a 28-node network using the CPLEX v.11.1 solver [15].

IV. DYNAMIC LSP CAPACITY RECONFIGURATION

In highly dynamic networks, such as OBS networks, the vol-ume of offered traffic may change abruptly and unexpectedly.In these scenarios, the previous offline VT dimensioning wouldbe inefficient only by itself. In the proposed GMPLS/OBSarchitecture, we also deploy an auxiliary mechanism to inducea dynamic character to the transport network.

The devised mechanism operates in a proactive manneravoiding LSPs to reach either full capacity, which wouldincrease the number of dropped bursts, or an inefficient use ofwavelengths. The decision to increase/decrease the number ofwavelengths associated to a LSP can be taken either on a localor end-to-end basis approach, spanning n-hops. Here, onlylocally based decisions spanning 1 single hop are considered,leaving for further work end-to-end alternatives. Specifically,the devised dynamic LSP capacity reconfiguration (DLCR)mechanism works as follows.

Each OBS node n ∈ V is responsible for monitoring the HPtraffic being offered by all LSPs supported on any of its outputlinks i, i = 1, ..deg(n), over a sliding temporal window Twith duration |T |. Note that by considering only one high-priority class, the LSPs share the same set of wavelengthsat each output link, facing the same wavelength occupancy.Therefore, the offered high-priority traffic load to an outputlink i in the node can be expressed as:

ρ(i) =

∑b∈B

tb

|T |(2)

where tb is the duration of the incoming burst b ∈ B, Bdenotes all the incoming HP bursts to be switched at node nwithin |T |.

At every monitoring interval, the OBS controller sends atrap message to its respective GMPLS controller reporting thecurrent HP traffic being offered to its output links. Upon recep-tion, the GMPLS controller is then responsible for detectingsudden traffic changes and triggering LSP reconfiguration ifrequired. Given ρ(i), it verifies whether the LSPs’ size at linki, Li (i.e. the number of wavelengths), is still appropriate. To

this end, it estimates the current HP traffic BLP using theErlang-B loss formula (B) and checks if the value remainsbelow the demanded QoS threshold Ω. For this, the followingassessment condition is verified:

eBLPHP = B(ρ(i), Li) < Ω ∀i (3)

The DLCR mechanism does not trigger the reconfigurationrequest before W consecutive windows with eBLPHP ≥ Ω(per each output link i). Such decision helps to maintain thestability of the system by remaining insensitive to short-termtraffic changes. If W is reached, however, the traffic peakis considered as significant and a 1-hop LSP expansion istriggered. To this goal, GMPLS computes the new set ofwavelengths for the LSPs on output link i to properly face themeasured HP traffic load. Such set of additional wavelengthsis given by: φ(i) = B−1(ρ(i),Ω) − Li where B−1(ρ(i),Ω)returns the number of wavelengths needed to satisfy Ω for thenew estimated traffic.

On the other hand, GMPLS can verify that the capacity ofthe LSP is over-dimensioned (i.e. φ(i) < 0). In such a case, agiven number of wavelengths may be released.

From the GMPLS control plane perspective, the rearrange-ment of those wavelengths assigned to the downstream link ofa given LSP is quite straightforward. In particular, the RSVP-TE module in the GMPLS controller records the informationof all configured LSPs in the node following the Path StateBlock (PSB) structure defined in [19][20]. For each LSP, theinformation contained in the PSB describes a similar structurethan the Path message that originally signaled it, namely, aSession, a Sender Template and an ERO object. As previouslydetailed in Figure 2, the Unnumbered interface ID sub-objectsin the Path message ERO has been here extended so that multi-ple wavelengths can be allocated at every LSP hop. Therefore,the allocation of additional wavelengths to a certain LSP isas easy as including in the specific Unnumbered interface IDsub-object those interface IDs associated to the wavelengths.Conversely, if some wavelengths already associated to the LSPwould have to be released, the associated interface IDs wouldbe removed from the specific Unnumbered interface ID sub-object. In this way, assuming that the LSP would have tobe released, that information in the PSB would be used todeallocate the resources supporting it.

V. PERFORMANCE ASSESSMENT

Simulations are performed to estimate the performance ofthe proposed QoS-aware GMPLS/OBS architecture under dif-ferent network scenarios and conditions. The objective is toprovide a validation of both i) planning strategies as the off-line VT dimensioning and of ii) operational strategies like theproactive reaction to unexpected traffic peaks. The simulationsare executed on the ad-hoc JAVOBS simulator developedin [21], here extended to implement the proposed architecture.

A. Simulation scenarios

Two reference network scenarios have been used in orderto claim for topology independence, namely, the 14-node

Fig. 3. HP burst losses with uniform (top) and non-uniform (bottom) traffic.

NSFNET and the 28-node EON network, where all linkssupport 32 bidirectional wavelengths at 10 Gbps. Regardingthe traffic characteristics, the [75%−25%] HP-BE traffic ratiois assumed for the whole set of simulations in all networkscenarios. All network nodes generate traffic following a Pois-son process, based on either uniform or non-uniform trafficmatrices. In particular, the HP traffic requires a BLP < 10−3

and is routed throughout a full-mesh VT of LSPs. Therefore,there is N∗(N−1) LSPs in each network scenario. Conversely,the BE traffic is routed using the Dijkstra’s Shortest Pathalgorithm. In the network, GMPLS/OBS nodes are equippedwith full wavelength conversion capabilities, a GMPLS controlnode, a traffic-engineered (TE) database to store LSP relatedinformation and an LSP occupancy monitoring system. Neitherdeflection or preemption mechanisms are initially consideredon the experiments.

B. Validation of the VT dimensioning model

Fig. 3 shows the BLP of all LSPs in the 14-node NSFNETnetwork, where a total load of 0.5 is offered according touniform (top) and non-uniform (bottom) traffic demands. Themain objective here is to validate the VT dimensioning model,ensuring that the requested QoS constraints in terms of BLPare finally meet in every established LSP. As seen, the experi-enced BLP in all individual LSPs is below the demanded BLP< 10−3 for both uniform and non-uniform traffic distributions.This not only validates the proposed dimensioning model,but also shows that it behaves independently of the trafficdistribution in the network.

C. Comparison of GMPLS/OBS against conventional OBS

Figs. 4 and 5 compare the proposed GMPLS/OBS archi-tecture against a conventional OBS architecture (convOBS).The results have been extracted from both NSFNET and

EON network scenarios, where a uniform traffic matrix isconsidered. In convOBS, deflection routing is applied for theHP class bursts as a QoS differentiation mechanism. From bothfigures, we observe an outstanding behavior of the proposedGMPLS/OBS architecture. As seen, the BLP of the HP trafficremains constant in the entire offered load range in bothNSFNET and EON network scenarios, being considerablybelow the demanded QoS threshold (10−3). In contrast, theBLP of the HP traffic grows exponentially with the offeredload to the network in convOBS, being soon over the BLPthreshold in the NSFNET network (from an offered load equalto 0.45) and never reaching it in the EON case. For example,differences around two orders of magnitude can be observedin the NSFNET network scenario for an offered load of 0.8.

Focusing on the BE traffic transmission, we have similarBLP values in both GMPLS/OBS and convOBS architectures,slightly higher in the former. The performance of the BE classcan be further improved when applying deflection routingor the burst preemption mechanism in the network. Suchscenarios are left for further study. However, we shall mentionthat the main concern here is to guarantee the QoS for HPtraffic, as agreed with potential OBS network clients throughService Level Agreements (SLAs).

D. Assessment of the dynamic LSP reconfiguration

In order to evaluate the dynamic LSP reconfiguration mech-anism proposed in Section IV, we enforce a peak of HP trafficof limited duration in a LSP. This leads to the followingactions: i) expansion of the LSP and ii) return to original LSPconfiguration. Fig. 6 shows a sample of the execution timeduring which we monitor the traffic occupancy of an outputlink of the NSFNET network. The QoS control is performed interms of HP burst loss probability (eBLP), which is estimatedevery T window with |T | = 5ms. From the figure, we observethat as soon as the traffic peak occurs, around t = 30ms(the offered load goes from 4,2 Er. to 8.5Er.), the GMPLSinstance detects it and remains on a holding stage duringW = 5 consecutive T windows (i.e. 25ms) with an eBLPabove the QoS level. Following the process as described inSection IV, those interface IDs of each additional wavelengthare included in the specific Unnumbered interface ID sub-object maintained in the PSB at the RSVP-TE module ofthe GMPLS controller. Once the PSB is updated, the OBScontroller is notified in order to update its forwarding table andthe LSPs acquire extended properties. As a result, the eBLPreturns to values below the QoS threshold. At approx. 400ms,however, the offered load is reduced back to its previous valueof 4.2Er (the traffic peak ends), so that the LSP becomes nowover-dimensioned for the current offered load. Therefore, inorder to achieve good resource usage, a counterpart processis triggered. After W = 5 consecutive T windows with extraallocated resources (i.e. wavelengths), the LSPs return to theirinitial configuration.

This DLCR mechanism induces a dynamic character(throughout GMPLS) to the proposed architecture to properlyhandle unexpected traffic demand surges. It is worth men-

Fig. 4. Network BLP: NSFNET Fig. 5. Network BLP: EON Fig. 6. Monitoring results

tioning, however, that long-term changes on the traffic matrixwould eventually require a reconfiguration of the VT.

VI. CONCLUSION

This paper introduced a GMPLS-enabled OBS network archi-tecture providing QoS-aware burst transport services. Aimingto assure the QoS requirements for the high-priority serviceclass, a MILP formulation was presented to compute an opti-mal virtual topology of burst LSPs over the OBS data plane.In this context, extensions to the standard GMPLS protocolsin order to set-up and dynamically reconfigure the computedvirtual topology were also proposed. Extensive simulationsresults highlighted that the proposed virtual topology designallows guaranteeing absolute BLP figures for HP class, evenin high load situations. Furthermore, the proposed dynamicburst LSP reconfiguration mechanisms yielded an efficientadaptation to those unexpected traffic surges, keeping the burstloss probability of the high-priority traffic below the requestedmaximum values.

ACKNOWLEDGMENT

The authors would like to thank their colleague FernandoAgraz for his valuable comments on the paper. This workhas been supported by the Portuguese Government Entity,Fundação para a Ciência e a Tecnologia (FCT), through thegrant SFRH / BD / 36950 / 2007, by the Spanish Ministryof Science and Innovation through the "COPERNic" project(TEC2009-13252), the Catalan Government under the contractSGR-1140, and the Polish Ministry of Science and HigherEducation under the contract 643/N-COST/2010/0..

REFERENCES

[1] S.J.B. Yoo, “Optical Packet and Burst Switching Technologies for theFuture Photonic Internet”, IEEE/OSA J. Lightw. Technol., Dec. 2006.

[2] C. Qiao, M. Yoo, “Optical burst switching - A new paradigm for anoptical Internet”, J. High Speed Networks, vol. 8, no. 1, Jan. 1999.

[3] J. Hwang, M. Yoo, “Congestion control algorithm based on predictionfor OBS networks”, in Proc. COIN-NGNCON 2006, Jul. 2006.

[4] J. Liu, N. Ansari, T. J. Ott, “FRR for latency reduction and QoSprovisioning in OBS networks”, IEEE J. Sel. Areas Commun., vol. 21,no. 7, pp. 1210-1210, Sep. 2003.

[5] Q. Chen, G. Mohan, K.-C. Chua, “Route optimization in optical burstswitched networks considering the streamline effect”, Computer Net-works, vol. 52, no. 10, pp. 2033-2044, Jul. 2008.

[6] Li Yang, G.N. Rouskas, “Adaptive path selection in OBS networks”,IEEE/OSA J. Lightw. Technol., vol. 24, no. 8, pp. 3002-3011, Aug. 2006.

[7] A. Pantaleo, M. Tornatore, A. Pattavina, C. Raffaelli, F. Callegati,“Dynamic service differentiation in OBS networks”, in Proc. Broadnets2007, pp. 37.39, Sep. 2004.

[8] F. Callegati, W. Cerroni, C. Raffaelli, P. Zaffoni, “Wavelength and timedomain exploitation for QoS management in optical packet switches”,Computer Networks, vol. 44, no. 4, pp. 569-582, Mar. 2004.

[9] A. Jajszczyk, “Automatically Switched Optical Networks: Benefits AndRequirements”, IEEE Commun. Mag., Feb. 2005.

[10] A. Farrel, I. Bryskin, “GMPLS: architecture and applications”, MorganKaufmann Series in Networking, 2006.

[11] L. Berger, “Generalized Multi-Protocol Label Switching (GMPLS) Sig-naling Resource ReserVation Protocol-Traffic Engineering (RSVP-TE)Extensions”, IETF RFC 3473, Jan. 2003.

[12] K. Kompella, Y. Rekhter, “Signaling Unnumbered Links in RSVP-TE”,IETF RFC 3477, Jan. 2003.

[13] M. Klinkowski et al., Graphs and Algorithms in Communication Net-works - Studies in Broadband, Optical, Wireless, and Ad Hoc Networks.pp. 165-181, (ed.) A. Koster and X. Muñoz, Springer-Verlag, 2009, ISBN978-3-642-02249-4.

[14] Z. Rosberg, H. L. Vu, M. Zukerman, J. White, “Performance Analysesof Optical Burst Switching Networks”, IEEE J. Sel. Areas in Commun.,vol. 21, no. 7, Sep. 2003, pp. 1187-1197.

[15] M. Klinkowski et al., “Virtual Topology Design in OBS Networks”, inProc. IEEE ICTON2010, Munich, Germany, Jun. 2010.

[16] R. Syski, Introduction to Congestion Theory in Telephone Systems,North-Holland, 1960.

[17] M. Minoux, Mathematical Programming: Theory and Algorithms, JohnWiley and Sons, 1986.

[18] M. Pióro and D. Medhi, Routing, Flow, and Capacity Design inCommunication and Computer Networks, Morgan Kaufmann, 2004.

[19] R. Braden, L. Zhang, “Resource ReSerVation Protocol (RSVP) – Version1 Message Processing Rules”, IETF RFC 2209, Sept. 1997.

[20] D. Awduche, L. Berger, D. Gan, T. Li, V. Srinivasan, G. Swallow,“RSVP-TE: Extensions to RSVP for LSP Tunnels”, IETF RFC 3209,Dec. 2001.

[21] O. Pedrola, M. Klinkowski, D. Careglio, J. Solé-Pareta, S. Rumley, C.Gaumier, “JAVOBS: a flexible simulator for OBS network architectures”,J. Networks, vol. 5, no. 2, pp. 256-264, Feb. 2010.